Transfer Learning Analysis of the Cox Mixture Model

ABSTRACT With the advancement of precision medicine, it is of great significance to identify potential patient types based on heterogeneous data and carry out risk assessment and measurement. The Cox mixture model is a highly significant tool for addressing such issues. However, as manifested in our numerical studies, the estimation of a Cox mixture model demands sufficient sample size; otherwise, it might lead to inferior classification accuracy and survival prediction. Motivated by the problem of insufficient sample sizes in estimating mixture models, this paper leverages a transfer learning framework and proposes a likelihood‐based approach incorporating an L‐1 penalty to improve the estimation efficiency and prediction accuracy. The asymptotic properties of the proposed estimators have been established. The numerical studies show that our proposed method possesses good performance under finite samples. Finally, two primary breast cancer datasets have been employed to illustrate the practical utility of our proposed method.